FiLM-Ensemble: Probabilistic Deep Learning via Feature-wise Linear Modulation
Authors: Mehmet Ozgur Turkoglu, Alexander Becker, Hüseyin Anil Gündüz, Mina Rezaei, Bernd Bischl, Rodrigo Caye Daudt, Stefano D'Aronco, Jan Wegner, Konrad Schindler
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, Fi LM-Ensemble outperforms other implicit ensemble methods, and it comes very close to the upper bound of an explicit ensemble of networks (sometimes even beating it), at a fraction of the memory cost. |
| Researcher Affiliation | Academia | Mehmet Ozgur Turkoglu ETH Zurich Alexander Becker ETH Zurich Hüseyin Anil Gündüz LMU Munich Mina Rezaei LMU Munich Bernd Bischl LMU Munich Rodrigo Caye Daudt ETH Zurich Stefano D Aronco ETH Zurich Jan Dirk Wegner ETH Zurich & University of Zurich Konrad Schindler ETH Zurich |
| Pseudocode | No | The paper describes the method using mathematical equations and textual explanations, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | We have implemented Fi LM-Ensemble in Py Torch and release the source code.1 [footnote] 1https://github.com/prs-eth/FILM-Ensemble |
| Open Datasets | Yes | We evaluate our proposed method on a diverse set of classification tasks, including popular image classification benchmarks, image-based medical diagnosis, and genome sequence analysis. CIFAR10 and CIFAR-100 [Krizhevsky, 2009] are widely used testbeds for image classification, and deep learning in general. [...] Retina Glaucoma Detection [Diaz-Pinto et al., 2019] is a real-world clinical dataset... REFUGE 2020 [Orlando et al., 2020]... 6m A Identification [Li et al., 2021], is a 1-dimensional sequential genome dataset. |
| Dataset Splits | No | For CIFAR-10 and CIFAR-100, the paper states: 'Each of the two datasets contains 60,000 images, of which 10,000 are reserved for testing.' This clearly defines the training and test sets (50,000 for training, 10,000 for testing), but no explicit validation set split is mentioned. |
| Hardware Specification | Yes | Measurements are done on an NVIDIA Ge Force GTX 1080 Ti. |
| Software Dependencies | No | The paper states: 'We have implemented Fi LM-Ensemble in Py Torch,' but it does not specify any version numbers for PyTorch or other software dependencies, which would be necessary for full reproducibility. |
| Experiment Setup | Yes | For all experiments, we optimize the model parameters with standard stochastic gradient descent, with momentum µ=0.9 and weight decay λ=0.0005 for regularization. We train for 200 epochs with batch size 128. The learning rate is initially set to 0.1 and decays according to a cosine annealing schedule [Loshchilov and Hutter, 2017]. The initialisation gain is set to ρ = 2 for all experiments, except for genome sequences (see Table 3), where ρ {4, 8, 16, 32}. |